SlidingWindow streaming: default rescue floors replace an unrelated opening word with a vocabulary term ("Hey" → "Codex")

[Alex-Wengg]

Surfaced while verifying #898 on the public #855 clip 01-validation-request-21.4s.wav (Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/Fixtures) with the vocabulary terms Codex, follow-up. Pre-existing: reproduces on main before #898.

The sliding-window streaming rescorer replaces an unrelated short word with the term through the spotter-rescue path at near-zero string similarity. Which word depends on the machine's decode, so this is not tied to "Hey":

• M5 Pro (local, GPU/ANE): Hey → Codex at the start of window 1.
• CI runner (macos-15 VM, CPU), fix(asr/vocab): tokenize in-code terms, rescore every streaming window, stop volatile-text loss (#851) #898 benchmark job on 230a5c1:

  'validate' vs 'Codex' (sim=0.12, span=1)
      -> REPLACE (vocab > original)
  [SPOTTER-RESCUE] 'validate' → 'Codex' spotter=-7.77, sim=0.12
  Final: Hey, before we go ... can you do me a favor and Codex
  

  and [MULTI] 'Hey, before' vs 'Codex' (sim=0.10) / 'to test with,' vs 'Codex' (sim=0.17) were considered.

In every case the real "codex" later in the sentence is left unchanged and batch decode keeps the original words. The streaming path uses VocabularyRescorer.Config.default; the unified path uses VocabularyBoostingSession.itnDefaultConfig (#702 floors 0.30 / 0.50) precisely to stop short-word rescue over-fire, and #862 recorded the same "rescue-floor WER tax" on the TDT path (recall up, WER 15.70 → 22.25 on .default).

Candidates:

• (a) Apply the CTC custom-vocabulary rescorer over-fires on short keyword terms since 0.14.8 (PR #634) #702 floors on the sliding-window streaming path by default. Directly rejects a 0.12 candidate and aligns the two streaming engines. Needs a recall/WER re-run on earnings22-kws before merging, since .default was kept on this path for brand-name recall.
• (b) Gate the rescue on the spotter detection's time range. Detections already map to overlapping word timestamps, but VocabularyRescorer+TokenRescoring falls back to the nearest word to the detection centre when no word overlaps, unconditionally — that is how a detection at ~9 s can land on Hey at 0.2 s or validate. Bound the fallback (e.g. within one word / marginSeconds) and log the detection interval so the trace shows which path fired.
• Also: the CLI's streaming path drops the rescue-floor flags (--vocab-spotter-min-sim, --vocab-spotter-min-sim-multi, --vocab-disable-spotter-rescue); only batch builds a VocabularyRescorer.Config from them. Pass them through runStreaming so the floors can be A/B'd from the CLI.

Regression guard once fixed: extend SlidingWindowVocabularyBoostingStreamingTests to assert the transcript still starts with "hey" and keeps "validate your findings with codex". Note that test currently cannot prove the rescorer acted (CI's decode already produced "follow-up" with zero replacements), so exposing spotter detections in ctcDetectedTerms even when no replacement is applied would give it something machine-independent to assert.

[Alex-Wengg]

Mechanism found with the debug trace: the CTC spotter reports a spurious low-score Codex detection in the first half second of every window (0.16–0.48 s), no word overlaps it, and the nearest-word fallback in wordIndices had no distance bound — so it rewrote whichever word was nearest (Hey 0.68 s away here, validate on the CI runner). #900 bounds the fallback to 0.5 s, surfaces spotter detections as ctcDetectedTerms even without a replacement, and passes the rescue-floor flags through the streaming CLI. Floors-by-default stays open pending a streaming recall/WER measurement.

[Alex-Wengg]

#900 is merged (b15ee10): the nearest-word fallback is bounded to 0.5 s, spotter detections are surfaced as ctcDetectedTerms on streaming updates (nil only when boosting is not configured), and transcribe --streaming honours the rescue-floor flags. Both observed corruptions (Hey → Codex locally, validate → Codex on the CI runner) came from a spurious detection in the first half second of each window snapping onto the nearest word; the bound stops that path. Leaving this open for the remaining decision — whether the #702 similarity floors (0.30 / 0.50) become the sliding-window streaming default — which needs a streaming recall/WER measurement first (the earnings benchmark drives the batch manager).